Google capped Meta's Gemini API capacity in March after demand outstripped supply

Google restricted how much Gemini model capacity Meta could purchase starting around March 2026, after Meta's requested usage outstripped what Google was able to supply, according to a Financial Times report.

What's new

Per the report, "Google, owned by Alphabet, told Meta around March it could not meet the full Gemini capacity the company had sought to purchase." Meta had been buying access to Gemini through Google's cloud and API services at a volume significantly above what Google could deliver, and the resulting cap has disrupted and delayed some of Meta's internal AI projects. In response, Meta reportedly directed employees to be more efficient with their AI token usage — the unit that measures compute consumption for AI workloads. As of late June 2026, the restrictions were still in place, and neither company had issued an on-the-record statement confirming the arrangement.

Context

Meta uses external frontier models, including Gemini, for internal work that spans safety processes such as scam and harmful-content detection, customer service, advertising tooling, and general internal workflows — alongside its own Llama and Muse Spark model families. Buying capacity from a rival frontier lab is a common practice even among companies that also build their own models, since internal training and inference needs frequently exceed available in-house compute. The episode surfaces amid a broader industry-wide compute crunch, with Google, Microsoft, Amazon, and others racing to expand data center capacity to meet demand from both their own products and enterprise customers.

Why it matters

The dispute is a concrete data point in the AI industry's compute-scarcity story: even Meta, one of the best-capitalized AI developers in the world, could not simply buy its way to more Gemini capacity when Google's supply was constrained. It also highlights an awkward dependency — Meta, a direct Gemini competitor through its own model lineup, relies on Google for parts of its AI infrastructure, giving Google leverage over a rival's internal operations when capacity runs tight. If compute constraints persist through 2026, expect more such disclosures of AI labs rationing access to each other's models and infrastructure.